Explainable machine learning for the prediction of Alzheimer’s disease-related cognitive impairment: a consensus feature selection approach

BACKGROUND: Early identification of Alzheimer's disease-related cognitive impairment remains challenging, and existing machine learning (ML) models often suffer from feature instability and limited interpretability. This study developed robust and explainable ML models using cerebrospinal fluid (CSF) biomarkers by systematically comparing sparsity-based (LASSO), importance-based (Boruta), and consensus feature selection strategies. METHODS: A publicly available cohort of 333 individuals (91 cognitively impaired, 242 cognitively normal) was analyzed. Data were split into training (70%) and independent test (30%) sets. Multiple classifiers, including Elastic Net-regularized logistic regression (LR), support vector machine (SVM), random forest, XGBoost, and Naive Bayes (NB), were trained using repeated 5-fold cross-validation (10 repetitions; 10 × 5-fold cross-validation) with class weighting. Model performance was evaluated using discrimination, calibration, and clinical utility metrics, and interpretability was assessed using SHAP. RESULTS: All models demonstrated strong discriminative performance on the test set (AUROC 0.861-0.958). LASSO-based models showed high specificity, Boruta-based models achieved higher sensitivity, and consensus-based models provided the most balanced performance. The consensus-LR and -SVM models achieved AUROC values of 0.954 and 0.951, respectively. Beyond discrimination, the consensus-LR model demonstrated good calibration and consistent net clinical benefit in decision curve analysis, analyses that remain relatively underreported in the Alzheimer's disease machine learning literature. SHAP analyses highlighted biologically plausible contributions from key biomarkers, including tau, Aβ42, NT-proBNP, pancreatic polypeptide, and IL-7. CONCLUSIONS: In summary, stable and interpretable ML models for Alzheimer's disease-related cognitive impairment can be developed using CSF-derived biomarkers obtained through lumbar puncture. The proposed consensus-based feature selection framework improves feature stability and model transparency, facilitating the discrimination between cognitively normal and impaired individuals and providing a foundation for future external validation studies.